Switchboard-1 Release 2
Linguistic Data Consortium (University of Pennsylvania)
Strangers chatting by phone about assigned topics in 1990, still a speech-recognition test.
About 2,400 two-sided telephone conversations among 543 American speakers, recorded by Texas Instruments for DARPA, each on a topic from a fixed list. It is the classic spontaneous-speech corpus for speech recognition, speaker identification, and the study of disfluencies and turn-taking, and many later annotation projects build on it.
Access
- Cost
- PaidFee for non-members; included in some LDC memberships.
- License
- ProprietaryLDC User Agreement for Non-Members
- Sign-up
- Purchase required
- How you get it
- Bulk downloadSPHERE (ulaw)
What's in it
- One row is
- One telephone conversation (two channels) with speaker and topic metadata
- Covers
- United States
- Years
- Recorded 1990-1991; LDC release 1993 and 1997
- Updated
- One-time release
- Size
- About 260 hours of 8 kHz audio
Quality
Well documented, with decades of follow-on transcripts and annotations; telephone-band audio and topic prompts limit realism.
- 8 kHz telephone audio
- Conversations follow assigned topics
- Paid license restricts redistribution
Connects to
- LibriSpeech ASR Corpus: Read-speech benchmark to compare against conversational errors
- Universal Dependencies Treebanks: Syntactically annotated text corpora, including spoken English treebanks
Questions it could answer
- How often do speakers say 'uh' versus 'um', and does it differ by sex or region?
- How much higher are speech-recognition error rates on Switchboard than on read speech?